Google Deepmind introduces the version to the device of its AI Gemini Robotics model, which allows it to work without an internet connection. The model of vision and language (VLA) has skillful possibilities similar to the one released in March, but Google says: “It is small and efficient enough to act directly on robot.”
The flagship model of Gemini robotics has been designed to support robots perform a wide range of physical tasks, even if it has not been specially trained on them. It enables robots to generalize recent situations and understand and respond to commands, as well as perform tasks that require excellent motor skills.
Karolina Parada, head of robotics at Google Deepmind, says The Verge The fact that the original model of Gemini robotics uses a hybrid approach, enabling it to operate the device on the device and on the cloud. But thanks to this model of devices, users can access offline function, which are almost as good as flagship functions.
The model on the device can perform several different tasks after removing from the box and can adapt to recent situations “from 50 to 100 demonstrations”, according to the parade. Google has trained this model of its robot Aloha, but the company was able to adapt it to various types of robots, such as the humanoid robot Apollo with AppTronik and the Robot FR3 FR3 Bi-sarm.
“The Hybrid Gemini Robotics model is still stronger, but in fact we are very surprised how strong this model is on the device,” says the parade. “I would think about it as a starting model or as a model of applications that simply have poor connectivity.” It can also be useful for companies with strict security requirements.
In addition, Google releases a software programming set (SDK) for a model on a device whose developers can utilize to evaluate and tune it-the first for one of the VLA Google Deepmind.
The Gemini robotics model on the device and its SDK will be available to a group of trusted testers, while Google is still working on minimizing safety hazards.